Chapter 5

Adaptive Self-tuning : A Fuzzy Logic Approach

5.1 Introduction

Fuzzy logic approach provides the most simple and efficient way of implementing a
feedback control mechanism in complex real-world applications [66] [67] like quality
assurance in food processing, pharmacy, speed control of locomotives, washing machines
etc. Application of fuzzy based control [68] is the most appropriate where the inputs and
outputs are generally expressed in linguistic terms. For instance, in a washing machine, the
weight of clothes being washed is expressed as light, medium and heavy. The weight of the
cloth is measured and is mapped through a membership function to indicate its degree of
belongingness to the three categories of weights. Depending on the degree of membership,
the speed of the washing machine is chosen again as slow, medium and fast. The output is
decided by a set of fuzzy rules and these have to be framed by the designer of the system
using his understanding of the system that needs to be fine-tuned. Fuzzy-based tuning
enables in building Policy-based [69] tuning systems.
89
 The DBA expresses the tuning needs in linguistic terms. For instance, he/she might
indicate that, if BHR is low then Buffer-Pool-size should be high depending on the current
value of the BHR. He/She might also use an adjective in addition to the linguistic terms.
For example, if the BHR may be expressed as very low instead of just low and the Buffer
pool size to be very high instead of just high. As discussed in the earlier chapters, DBMS
being a complex software system with its own set of tuning parameters having diverse set
of values, application fuzzy is the most appropriate choice for implementing feedback-loop
based tuning-action.

5.1.1 Fuzzy Inference Systems

The fuzzy rule-based systems, fuzzy expert systems are known as Fuzzy Inference
Systems(FIS). The major function of this unit is decision making based on fuzzy rules that are
of “IF THEN” type. These constructs might make use of conjunction operators like OR, AND,
NOT while framing the fuzzy rules. The input to FIS may be fuzzy or crisp but the output will
always be fuzzy. However, when FIS is used in a control system, the output must be crisp.
Hence, the output from the FIS must be defuzzified to get the values in the crisp form. The
working methodology of FIS is as under :
1. Inputs are fuzzified using different membership functions that include, triangular,
trapezoidal and gauassian membership functions.
2. The rules are framed and stored in a database and this forms the knowledgebase for the
FIS.
3. Finally output is defuzzified using appropriate steps.

There are two important types of FIS models, namely Mamdani model and Sugeno
model[ 103]. In the former the output from FIS will be a fuzzy set. Hence, defuzzification is a
must, whereas in the latter, the output is determined by a polynomial in the input variables and
is first order. The output from Sugeno model will always be crisp and hence no defuzzification
step is required. Though Sugeno model is computationally efficient and adaptive in nature,
Mamdani model is preferred as it has wide acceptance, well suited for human input/output and
is more intuitive in nature [61]. Hence, in the adaptive fuzzy tuning technique, Mamdani model
is employed.

90
5.1.2 Membership functions for fuzzication

Membership functions are used to transform crisp input values of the real world
application to fuzzy sets irrespective of the elements of the inputs, which are discrete or
continuous.

Figure 5.1 The Fuzzification of User load with Gaussian membership function

Several membership functions have been proposed by experts in the field that include,
triangular, trapezoidal, Gaussian etc. The membership functions are characterized by three
important features.

1. Core : Core of a membership function is a region of the input where the membership
value is exactly equal to 1. It means the inputs which fall in this category have complete
membership.
2. Support : Support of a membership function is a set of input-values where the
membership value is greater than zero.
3. Boundary : This part of the support where the elements that have partial membership
that are non-zero.

The following notation is used to indicate the membership function values for the members of
the input universe. If input X=[x1,x2,……..xn] then the membership value of the fuzzy set A is
defined as

A = { µ(xi), xi | X } ---------------------------------------------(5.1)

91
The membership values can be assigned using either intuition, inference or inductive reasoning,
rank ordering, neural network methods. For all the input and output variables in the proposed
self-tuning system, guassian membership function[61] was used as it has smooth edges and
hence there would be no abrupt changes in the estimated tuning parameter values.

Figure 5.2 Fuzzification of DB_Cache tuning parameter

5.2 Fuzzy Controlled tuning

Fuzzy logic is the most suitable choice for many control applications for the fact that
fuzzy control systems are robust and can be tweaked easily to improve the system performance
dramatically and most importantly they are much simpler in design to implement. Moreover,
there is no need to measure the rate of change of the input parameters and the number of inputs
and outputs are not limited to small number. In the proposed setup the input sensor parameters
Buffer-Hit-Ratio(BHR), User-load(N) and Database Size(DBS) are extracted from the DBMS
and are fuzzified before being fed to the Fuzzy Inference System(FIS).

5.2.1 Fuzzy Controlled Tuning Architecture

Figure 5.3 shows the generic autonomic computing architecture modified to implement
self-tuning architecture for the DBMS wherein the FIS forms an important component of the
entire self-tuning system. The FIS takes the sensor inputs namely the BHR, Userload N and
Database Size as input to implement, Analyze and Plan modules of the MAPE based autonomic
computing system. The inputs have to be fuzzified using appropriate membership functions
[62] before being fed to the FIS. Similarly the output estimated by the FIS using fuzzy rules

92
from the Knowledge-base must be defuzzified before being used by the Execute module to
fine-tune the DBMS.

Knowledge-Base : Fuzzy Rules : Training Dataset

Fuzzy Inference system to estimate

the tuning parameters
Aggregator
Monitor Neural Network to estimate Execute
the tuning parameters

Sensor Managed Resource Touch Point Effector

Database Management DBMS Memory

System Components

End User App.

Database

Figure 5.3 The General Self-tuning architecture using Fuzzy Inference System

The FIS uses the fuzzy rules that are in the form “IF Then” to estimate the three
tuning parameters namely, DB_Cache, Shared_pool and Large_pool. These three estimated
parameters must be defuzzified to generate the crisp values before being used by the tuning
module to fine-tune the system.

The defuzzification uses a mean max method to generate crisp output values. Figure 5.4
shows the complete fuzzy based self-tuning setup. The extraction module uses SQL commands
and scripts to determine the BHR, User-load N and Database size DBS and feeds to the
fuzzifier module.

5.2.2 Fuzzy sub-system implementation

Figure 5.4 shows the architecture of fuzzy controlled self-tuning system. The objective
of this system is to analyze the DBMS, by proactively monitoring the performance indicators
like buffer-hit-ratio, number of active processes and the database size that are showing signs of
rapid growth and initiate control measure using fuzzy control. The inputs are BHR, N and DBS
that are crisp and hence, have to be fuzzified before being acted upon by the FIS.

93
 New Buffer_Cache_Size

New Shared_Pool_Size
Tuner
Module
New Large_Pool_size

DBMS

N
Performance Fuzzyfier
BHR Indicator Module Fuzzy De-Fuzzyfier
Extraction Controller Module
DBS Module Module

DBA

DATABAS Fuzzy Rules

Tuning Rules

Figure 5.4.. Fuzzy controlled

contr based self-tuning architecture

The FIS uses the fuzzy rules stored in the knowledgebase to decide on the values of the
output variables of the fuzzy system.

Figure 5.5 Computation of Output parameter of the Fuzzy inference

inferen system

In the fuzzy based tuning system the output variables are the tuning parameters
themselves. The output fuzzy tuning values have to be defuzzified before being applied to the
database system. The defuzzification uses centroid method to convert the
the fuzzy output
variables to their crisp counterparts. Figure 5.5 shows the evaluation of the tuning parameter as
the inputs vary using fuzzy rules.
94
5.2.3 Fuzzy Control Rules

Fuzzy control rules are based on general logical reasoning and based on programming
language constructs namely the IF THEN control construct. Linguistic terms are used to
describe the extent of tuning required.
Table 5.1 Fuzzy Inference Rules

Rule 1 IF BHR is Best AND the user load is Low and DBSize is High THEN set BCS to
low.
Rule 2 IF BHR is Best AND the User load is Medium and DBSize is High THEN set BCS
to Moderate.
Rule 3 IF BHR is Good AND the User load is High and DBSize is High THEN set BCS to
High
Rule 4 IF BHR is Low and the User load is High and DBS is high THEN set BCS to Very
high
Rule 5 IF BHR is High and User load is Less and DBS is Small THEN set BCS to Low
Rule 6 IF BHR is Medium and and Userload is Medium and DBS is Medium THEN set
BCS to Medium
Rule 7 IF BHR is Low and Userload is Medium and DBS is Large THEN set DCS to High
Rule 8 IF BHR is low and User load is Very High and DBS is Large THEN set DCS to
Very High
Rule 9 IF BHR is High and Userload is High and DBS is Large THEN set SHP to Low
Rule 10 IF BHR is High and Userload is High and DBS is Large THEN set LRP to Medium
Rule 11 IF BHR is Low and Userload is High and DBS is Medium THEN set SHP to
Medium
Rule 12 IF BHR is Low and Userload is High and DBS is Medium THEN set SHP to High
Rule 13 IF BHR is low Userload is High and DBS is Small THEN set SHP to low

Buffer Cache Size, DBS

(BCS

Database size, BHR


Buffer Hit Ratio)

These rules must be carefully framed and most of the time require modifications till desired
results are obtained. For instance, to find the new buffer caches size, some of the fuzzy rules
could be formed as shown in Table 5.1. The values in figures 5.1 and 5.2 were fixed based on
the user-load range of 1-100 and tuning parameter range of 0-1400MB respectively. The values
in figure 5.5 were determined by the fuzzy-rules, the membership function used and the

95
defuzzification method employed. In this setup, Gaussian membership function was employed
and mean of max method of defuzzification has been used.

5.2.4 Result and Analysis

The adaptive self-tuning method based on Fuzzy-control has been validated using
standard workloads namely, TPC-C and TPC-H for its effectiveness in fine tuning the DBMS.
Figure 5.6 shows the results wherein, the Fuzzy-control response-time is compared with auto-
tuning feature of Oracle 10g under TP-C workload (SF=2). Due to rigid rule-based approach,
the Fuzzy inference system slightly underestimates the tuning parameters values for the user-
range(2-10) and hence, in the initial portion of the graph, the response-time is slightly higher
than that of the auto-tuning method. However, for user-load beyond 50, the response-time is
steady. The performance of Fuzzy-controlled tuning shows an improvement of 20.8% in
response-time as compared to Auto-tuning feature of Oracle 10g.

250

200

150
Fuzzy_Tun
Response time(mescs.)

Auto-Tun
100

50

0
2
6
10
14
18
22
26
30
34
38
42
46
50
54
58
62
66
70
74
78
82
86
90
94
98

User Load N

Figure 5.6 : Response-time v/s Userload using Fuzzy based self-tuning

The fuzzy controlled tuning technique is also validated under TPC-E workload, SF=1
and the above graph in figure 5.7 shows and improvement of 28.7% over the built-in auto-
tuning feature of Oracle 10g.

96
 Figure
5.7 : Response-time V/s User load : TPC-E Workload, SF=1

5.2.5 Fuzzy Controlled and Neural Network Controlled Tuning : A Comparison

Neural-Network-based tuning method shows much better performance over the entire
user-load range of 2-100 as compared to fuzzy approach. However, it is much harder to
implement Neural-network-based tuning system, as the construction of the training data- set is
both time consuming and tedious. Further, the training data-set has to be framed for every kind
of workload and scaling factor. On the other hand, Fuzzy based tuning technique is much easier
to implement as it is simpler to frame fuzzy rules without the need to know the tuning impact of
each tuning parameter on the response-time. However, it is important to know the range of each
tuning parameter over which tuning is effective. Though, this method is simpler, it does not
provide an improved response-time for all workload types over the entire range of the user-
load.

Summary

In this chapter, a fuzzy-rule based adaptive tuning method has been presented. The
input variables namely user-load, buffer-hit-ratio and the database size are fuzzified before
being used by the Fuzzy Inference System. The FIS then uses the knowledgebase that is in the
form of fuzzy rules to estimate the output fuzzy variables. The defuzzification generates the
crisp tuning variables that are used by the tuner module to effect changes to the dynamic tuning
parameters of DBMS. The method has been validated under two workload types namely TPC-
C and TPC-E, and the results show an improvement of 18.52 and 28.7% improvement in

97
performance respectively. However, with TPC-C the performance is not at par with the Neural
network based auto-tuning of the DBMS. This may be due to rigid rule-based nature of the FIS
in estimating the tuning parameter values.

98